Unveiling spatial correlation between cell surface targets and lymphocytes via artificial intelligence (AI)-powered quantitative immunohistochemistry (IHC) analysis

Background With increasing emphasis on the relationship of lymphocyte distribution with cell surface targets in immunotherapy, a standardized and reproducible analytical Methods is essential. We propose an AI-powered Methods to assess the relationship of lymphocyte distribution with 74 membrane-specific targets in development by quantifying the density of lymphocytes and target expression.

Methods A total of 47,591 cancer and normal tissue IHC images from Human Protein Atlas were analyzed on 74 target genes. The AI model trained on pathologists-annotated whole-slide images (WSIs) analyzed intra-tumoral/extra-tumoral infiltrating lymphocyte (iTIL/exTIL) densities and tumor proportion score (TPS) of surface targets using the IHC WSI. iTIL, exTIL densities were compared between the TPS ≥1% and TPS <1% samples according to cancer types (pan-cancer and each cancer types respectively). Positive/negative correlation with TIL with protein expression are defined as statistically significant (p <0.05) 2-fold increase/decrease in average TIL in TPS ≥1% samples. Targets not meeting the above criteria were considered non-significantly different.

Results The IHC analyzer examined 174M cells including 109M cancer cells in 34 cancer types. Among 74 targets analyzed, mean TPS of all targets was 17.4%(± 16.4%), with 49.0% of WSIs (12,939) classified as TPS ≥ 1% and 51.0% (13,443) classified as TPS <1%. In pan-cancer analysis, TIL density was mostly either decreased (23.0%) or not significantly changed (74.3%) in the TPS ≥ 1% group, but only 2 (2.7%) targets including PD-L1 were positively correlated with total TIL (iTIL+exTIL). Moreover, iTIL significantly increased in the tumor tissues with PD-L1 TPS ≥ 1% in pan-cancer analysis (mean iTIL 510.2/mm2 vs 237.5/mm2, p <0.001). In subgroup analysis by cancer type, a positive correlation is observed in 24 pairs of target-cancer type. Among these FGFR4 showed positive correlation with iTIL in colorectal cancer (mean iTIL 88.2/mm2 vs 8.1/mm2, p = 0.011), non-squamous lung cancer (75.0/mm2 vs 25.9/mm2, p = 0.015) and uterine carcinoma (225.4/mm2 vs 31.7/mm2, p = 0.003) but not in other cancer types that highly express FGFR4 such as of hepatocellular carcinoma or squamous lung cancer.

Conclusions This correlative pipeline reveals a lack of pan-cancer correlation for cell surface targets and TIL infiltration, with the exception of PD-L1. However, expression of FGFR4 was linked to TIL presence in select cancer types which suggests the possible utility of a FGFR4 T cell engager bi-specific.